slides

What Difference a Group
Makes
Web-Based Recommendations for
Interrelated Users
Anthony Jameson and Barry Smyth
Tomek Loboda
Additional Papers
• Adaptive Radio: Achieving Consensus Using
Negative Preferences
By  Dennis L. Chao, Justin Balthrop, Stephanie Forrest
In  Proceedings of the 2005 International ACM SIGGROUP
Conference on Supporting Group Work, New York, 2005.
At 
http://www.cs.unm.edu/~dlchao/papers/chao05group.pdf
• Explaining Collaborative Filtering
Recommendations
By  Jonathan L. Herlocker, Joseph A. Konstan, and John Riedl
In  Proceedings of the ACM 2000 Conference on Computer
Supported Cooperative Work, December 2-6, 2000.
At  http://www.grouplens.org/papers/pdf/explain-CSCW.pdf
Outline
•
Introduction
1.
2.
3.
4.
Preferences Specification
Preferences Aggregation
Explaining Recommendations
Final Decision
•
Conclusions
Introduction
I
Scenarios
• Group of friends going on vacation
together
• A family selecting movie or TV show to
watch together
• A group of colleagues choosing a
restaurant for an evening out
• Keyword  together
I
Issues: Phase 1
• Members specify their preference
• It may be desirable for members to
examine each other’s preference
specification
• What benefits and drawbacks can such
examination have and how can it be
supported by the system?
I
Issues: Phase 2
• The system generates
recommendations
• Some procedures for aggregating
preferences must be applied
• What conditions to such aggregation
procedures have to be fulfilled and
what kind of procedures tend to fulfill
them?
I
Issues: Phase 3
• The system presents recommendations
to the members
• The (possibly different) suitability of a
solution for the individual members
becomes an important aspect of
solution
• How can relevant information about
suitability for individual members be
presented effectively?
I
Issues: Phase 4
• Members decide which
recommendation (if any) to accept
• The final decision is not necessarily
made by a single person – negotiation
may be required
• How can the system support the
process of arriving at a final decision, in
particular when members cannot
engage in a face-to-face discussion?
I
vs Collaborative Filtering
• CF model groups of users…
• …using similarity measure, i.e. shared:
– Preferences
– Rating patterns
– etc.
• GR model groups of users…
• …defined by a social context:
– Potentially much less similarities
I
Systems
•
•
•
•
•
Let’s Browse  browsing the Web
PolyLens  movie recommender
Intrigue  tour guide assistant
MusicFX  automatic music selection
Travel Decision Forum  vacation
planner
• I-Spy  Web search engine
• Adaptive Radio  music broadcast in a
shared environment
1
Preferences
Specification
PS
Methods
• Implicit – whenever possible
• Explicit – sometimes unavoidable
• MusicFX  rating 91 genres
• Travel Decision Forum  rating variety
of attributes of vacations destinations
• Adaptive Radio  censoring disliked
songs
• I-Spy  selecting a link
PS
Sharing Preferences
1. Saving effort
2. Learning from other members
“I can’t go hiking, because of an injury”
•
Travel Decision Forum  copy + edit
PS
Travel Decision Forum

PS
Sharing Preferences
• Taking into account attitudes and
anticipated behavior of other members
• Encouraging assimilation to facilitate
the reaching of agreement
2
Preferences
Aggregation
PA
Individual Models
1. For each candidate c
– For each member m predict the rating rcm
– Compute an aggregate rating Rc
2. Recommend the set of candidates
with the highest predicted ratings Rc
•
Rc = max rcm
PA
Group Model
1. Compute an aggregate preference
model M that represents the
preferences of the group as a whole.
2. For each candidate c use M to predict
the rating Rc for the group as a whole.
3. Recommend the set of candidates
with the highest predicted ratings Rc.
PA
Group Model: Details
• Preferences defined and negotiated
once
• Privacy issue avoided
• Recommending (accurately or not) a
candidate for which the predicted
rating of each individual member was
low
PA
Goals and Procedures
•
•
•
•
Maximizing average satisfaction
Minimizing misery
Ensuring some degree of fairness
Discouraging manipulation of the
recommendation mechanism (…)
• Ensuring some degree of
comprehensibility
• Treating different group members
differently (where appropriate)
PA
Counteracting Manipulation
• Not showing preferences of other’s
before specifying own ones:
– Guessing (at least roughly)
– Advantages of showing them
• Manipulation can most likely happen…
• …with explicit preference specification
as an input
PA
Counteracting Manipulation
Solution…
• …explicit model with trust factor
• I-Spy:
– Clicking on a link promotes it – easy to
observe and abuse
– Each link selection action is evaluated for
reliability
– No promotion unless a threshold is reached
PA
Counteracting Manipulation
• Inherently nonmanipulable
mechanism?
• Automatic generation?
• Travel Decision Forum:
– Median of the individual preferences used
as a preference of the group as a whole
– Automatic generation on the fly
PA
Right Procedure – Levels
1. By designers – before deployment:
– System’s goals/assumptions/context
driven
– PolyLens  small groups  “least misery”
aggregation function
– Avoiding manipulation not always
necessary:
•
Purchasing decisions as input
PA
Right Procedure – Levels
2. By users – aggregation function
selection:
– Before any recommendations are made
– During an iterative process of requesting
recommendations
– Intrigue  different weights for
subgroups
– Travel Decision Forum  variety of
aggregation mechanisms
PA
Right Procedure – Levels
3. By users – specific recommendation
consideration:
– User compiles the recommendations with
the goal in mind before making the final
decision
– The system should take those goals into
account too to ensure that the set of
candidates includes one or more highly
suitable option
PA
Multiple Decisions
• Larger set of decisions at the same time
or in succession
• Intrigue  several sights to visit
• Let’s Browse  number of web pages in
the course of a given session
• Local vs. Global:
– L: none of the members satisfied
– G: each member satisfied some of the time
• Decisions as a package
3
Explaining
Recommendations
ER
Motivation
• Black box does not provide insight into:
– How the recommendations were arrived at
– How attractive they are for each individual
member
• Explanations:
– Provide transparency
– Helps to detect sources of errors
• Important especially in high-risk
domains
ER
Benefits for the User
• Justification – how much confidence
will I place in that recommendation
• User Involvement – user adds their
knowledge and inference skills
• Education – better understanding the
strengths and limitations
• Acceptance – as a decision support tool
ER
Let’s Browse
ER
PolyLens
G
I
ER
Intrigue
ER
I-Spy
• How other community members have
dealt with a given page
• Information offered:
– Related queries
– Quantitative and temporal information
ER
Travel Decision Forum
ER
Travel Decision Forum
• Simulated reaction effects:
– Increased awareness of other members’
point of view
– Overcoming the natural tendency to focus
on one’s own evaluations
4
Final Decision
FD
Considerations
• The decision is not made by a single
person and therefore…
• …extensive debate and negotiations
may be required
• Can members communicate easily?
FD
Existing Approach 1
1. Autonomous translation of most
highly rated solutions into actions
•
•
MusicFX  changes the channel
Adaptive Radio  changes the song
•
Good when a quick decision is needed
FD
Existing Approach 2
2. One group member is responsible for
making the final decision
•
•
•
Let’s Browse  one person is
controlling the pointing device
Intrigue  tourist guide is deciding
what tour should be taken
Makes the system design simpler
FD
Existing Approach 3
3. Conventional conversation (face-toface or by phone) as a medium
•
PolyLens  members can call/IM each
other
•
Makes the system design simpler
FD
Existing Approach 4
4. Built-in communication support
•
Travel Decision Forum  avatars can
be granted a certain amount of
authority to accept proposals during
interactions with other members
•
Makes the system design difficult
FD
Possible Extensions
• Thresholds of acceptance
• Voting
• Isn’t voting itself recommending?
– Seeing or not seeing votes of others
– Counting and weighting votes
– Presenting the results of voting
– How to arrive at the final decision
• Yes, but in a much simpler context…
• …which means it is better defined
Conclusions!
C
The 4 Phases
1.
2.
3.
4.
Preferences Specification
Preferences Aggregation
Explaining Recommendations
Final Decision
C
Group Recommenders
• Only few systems
• Small subset of possible
recommendation techniques
• Limited number of application domains
• Different superficially only
• Context dependant
C
Other Systems
• What happens if…
• …we have groups of cooperating,
information seeking user?
• We can see how would the 4 phases be
applied
• For instance, adaptive navigation
support
C
I-Spy Demo?
Thank you!